Skip to main content
Sometimes you don’t want a permanently-synced projection — you want the data moved once: seed a new search index, populate a staging environment, or lift a dataset into a serving store and be done.
A dedicated one-shot mode (bootstrap and exit) is a planned engine capability and is not in the released binary yet. The pattern below achieves a one-time copy with the shipped engine today.

​
The pattern today

A one-time copy is the zero-downtime migration flow, stopped early:
  1. Run the pipeline with bootstrap: mode: snapshot. The engine captures a cursor, streams every existing row through your projection into the target at bulk speed, then switches to tailing.
  2. Wait for the tail to catch up (cursor age near zero) — this guarantees rows written during the copy were also delivered, which a plain dump script cannot promise.
  3. Stop the engine. The target now holds a complete, consistent copy as of the moment you stopped.
Because delivery is idempotent and the cursor is durable, an interrupted run resumes exactly where it left off — restart the engine and the copy completes rather than starting over.

​
Verifying the copy

Before declaring the migration done:
  • Compare source row count with target document count (deterministic ids mean one row → one document; no duplicates to distort the count).
  • Spot-check recently-written rows — anything committed before you stopped the engine must be present.
  • The engine’s delivery counters (events_delivered_total, dlq 0) confirm nothing was routed to the dead-letter queue.

​
What the planned one-shot mode adds

The dedicated mode will remove step 3’s judgement call: the engine will bootstrap, drain the change stream to the captured high-water mark, report parity, and exit on its own — making it scriptable in CI and batch contexts. Watch the release notes for availability.